henhaobio
Sign in

bioRxivPlant BiologyPreprint

BOTANIC-1: a series of long-context plant genomic foundation models in the agentic era

Amélie BarozetiD, Vincent CabeliiD, Jean Ogier du TerrailiD, Alexey RukhovichiD, Thomas JanssooneiD, Gary KlajeriD, Zeinab SheikhitarghiiD, Gregory AndrewsiD, Cyril Véran, Léonard Strouk

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The development of climate-resilient crops would be greatly accelerated by models able to reason directly over plant genomic sequences and to pinpoint trait-associated regions or loci. Anticipating the impact of DNA base changes (variants) remains challenging, and understanding regulatory mechanisms is still an active area of research. Through self-supervised training on unannotated genomic data, genomic language models (gLMs) can learn DNA syntax and grammar that go beyond current annotations, thus complementing standard bioinformatics analyses that rely on rules established by decades of genomics research. Here we present our agent-powered Model Factory and its first outputs: the Botanic1 family of gLMs designed for plant research, which operates reliably on sequences from hundreds of base pairs up to 128 kbp. These models outperform all generalist and plant-specific gLMs (as well as specialised baselines) on one of the largest sets of plant genomics evaluation tasks reported to date, at a much smaller budget than concurrent models. Mechanistic interpretability analysis identifies features associated with biologically meaningful sequence properties including coding region boundaries and splice site motifs, demonstrating that these models are a source of biological insight beyond their benchmark performance. Finally, because a gLM only becomes practically useful when embedded in a broader workflow, we integrate Botanic1 as a specialised genomic layer callable by a generalist large language model (LLM) agent, illustrating how such hybrid systems could accelerate plant biology research. To support the plant genomics research community, we release the four Botanic1 models, their pre-training corpus and the trained sparse autoencoder for research use at https://huggingface.co/spaces/living-models/botanic1-report.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (23 Sept 2026) · CC BY 4.0

Bidirectional Mamba2 plant genomic language models (318M to 3.2B parameters) trained on 320 land-plant genomes top a 22-task frozen-model aggregate; a small variant reaches 131,072 bp context.

Pinpointing causal crop variants is slowed by linkage disequilibrium, sparse plant annotation and poorly mapped regulation, while earlier plant gLMs trained on fewer than 100 species with limited systematic data curation. The work applies agent-run, seed-replicated data ablations to a 320-genome land-plant corpus, releases four models, and tests them from frozen probes to zero-shot causal-variant ranking inside an LLM agent.

Key results

  1. ShownThe four Botanic1 sizes occupy the top four places on a species-balanced aggregate of 22 frozen-model plant tasks, ahead of generalist and plant-specific gLMs many times larger, though the lead of Botanic1-S over PlantCAD2-L is not statistically significant.

    Aggregate: Botanic1-S 0.758, -M 0.764, -L 0.768, -XL 0.770 vs PlantCAD2-L 0.752, Carbon-8B 0.723, Evo 2 7B 0.716; L and XL margins over PlantCAD2-L significant under both bootstrap tests (+0.016, +0.018), S margin +0.006 not; Botanic1-S used 318M vs 694M parameters and 314.6B vs 4T tokens. · 1 Introduction; Best overall performance and efficient training; Plant gLM benchmarks measure different aspects…; Results

  2. ShownZero-shot likelihood-ratio scores rank experimentally validated causal SNPs near the top of their loci more often than competing gLMs, and as a callable tool Botanic1-L lets a small LLM agent rank the melon CmEIN3 causal variant first.

    545 loci, 14 species: Botanic1-XL recall within top 1% 0.499 vs 0.446 PlantCAD2-L and 0.429 PlantCaduceus-l32 (paired CIs exclude zero). Melon locus, 3,061 candidates, 20 runs: Gemma 4 E4B with Botanic1-L reached Recall@1 of 1, vs Recall@20 of 0.35 with expert-guided conventional tools and 0 with no tools. · 2.5.1.2 Zero-shot causal variant prioritisation; 2.5.2 Botanic1 is the foundational building…

  3. ShownFine-tuned Botanic1-S beats specialised and gLM baselines: it surpasses ChromBPNet on base-resolution ATAC-seq in Arabidopsis and maize (a gain absent when the architecture is trained from random weights) and leads AgroNT and PlantCAD2-L on promoter and terminator strength.

    Peak count Pearson 0.750 to 0.765 vs 0.666 for ChromBPNet in Arabidopsis and 0.847 to 0.857 vs 0.809 in maize, paired-bootstrap CIs above zero for every metric; terminator Pearson r 0.861 vs 0.834 (AgroNT) and 0.790 (PlantCAD2-L); promoter 0.896 and 0.874 vs 0.864 and 0.831 (AgroNT). · 2.3.2.1 Parameter-efficient adaptation of…; 2.3.2.2 Pre-training drives Botanic1’s…; Botanic1 outperforms ChromBPNet thanks to…

  4. ShownStaged continued pre-training of a Botanic1-S-sized backbone to 131,072 bp windows keeps short-context scores stable while letting the model exploit distal sequence for perplexity, key-value retrieval and, with peak-centred pooling, chromatin accessibility prediction.

    Short-task Sbal stays near 0.73 through the curriculum; nested-context PPPL optimum improves from 2.58 (Botanic1-S0) to 2.40 (Botanic1-S4); retrieval score rises from 0.56 to 0.83 vs 0.56 for PlantCAD2 at 694M; peak-centred pooling lifts median ACR AUPRC to 0.77 at 32,000 bp vs a 0.74 full-window peak. · 2.2.3 Context-extending Botanic1-S to 131 kbp…; 2.3.3.1 Botanic1-S benefits from…; 2.3.3.2 Context-extension improves Botanic1-S…; 2.3.3.3 Chromatin accessibility prediction…

  5. ShownAn SAE on Botanic1-M layer 35 yields single features detecting splice donors, acceptors and introns across four species, plus codon-position features; two donor features track splice-site usage and disagree with a TAIR10 donor at AT1G65170 that TAIR12 later moved to the predicted site.

    Donor feature f4131 MCC 0.871 and acceptor f586 MCC 0.873 (0.821 to 0.937 across Arabidopsis, rice, tomato, maize); two-feature tree separates constitutive, intermediate-usage and decoy donors at balanced accuracy 0.841 vs 0.643 for PWM/Markov scores; activation vs usage Spearman ρ 0.753 and 0.759. · 2.4.1 Base and region-level features; 2.4.2 A splice feature anticipates a change in…; 2.4.3 Alternative splice sites are identifiable…

Methods, data and reproducibility

Single-nucleotide masked language modelling trains bidirectional Mamba2 encoders (318M to 3.2B) on 8,192-token windows from 320 land-plant genomes, with data choices fixed by two-seed ablations on a species-balanced metric; models are assessed by frozen probes, zero-shot LLR, fine-tuning, context extension, sparse autoencoders and an LLM agent.

  • Botanic1 pre-training corpus: functional-balanced, MinHash-deduplicated 8 kbp windows from Ensembl Plants release 62 and NCBI assemblies · land plants (Embryophyta), 320 contributing species · n = 7,207,506 windows, 59.0 Gbp; 326 genomes
  • Plant Genomic Benchmark (PGB) and PlantCaduceus dataset suites, plus GPN-derived genomic region tasks and PGB fine-tuning tasks (STARR-seq strength, poly(A), lncRNA) · multiple plant species (e.g. A. thaliana, O. sativa, Z. mays, sorghum) · n = 22 metrics in Sbal; 34 per-task-species metrics overall
  • New zero-shot benchmark of loci centred on experimentally validated causal SNPs, each scored against all documented SNPs in a 100 kbp region · 14 plant species · n = 545 loci (29-study subset used in Sbal)
  • Population SNP panels from 1001 Genomes, CicerSeq (CDC Frontier v1) and RiceVarMap2.0 for MAF, missense/synonymous and phyloP analyses · Arabidopsis thaliana, Cicer arietinum, Oryza sativa · n = 6,491,238 / 880,345 / 6,494,573 SNPs
  • Bulk segregant analysis chromosome 2 sex-determination locus (Rashid et al.) used for the retrospective agent experiment · Cucumis melo · n = 3,061 candidate variants
  • Base-resolution ATAC-seq tracks and peak sets processed through a plant-specific pipeline, scored on held-out chromosomes · Arabidopsis thaliana; Zea mays · n = 4,646 (A. thaliana) and 23,220 (maize) held-out peaks
  • Cell-type-specific accessible chromatin region dataset from Zhai et al. for context-length experiments · Zea mays · n = 92 cell types
  • PastDB alternative 5' splice-site usage from RNA-seq, used to define constitutive, intermediate-usage and decoy donors · Arabidopsis thaliana · n = 885 sites per class; 3,254 annotated donors
Methods & tools
BiMamba2 (bidirectional Mamba2) backbone, Meta Lingua pre-training library (forked), sourmash (k = 31, scaled 5,000) for ANI-based species distances, MinHash near-duplicate window filter, BatchTopK sparse autoencoder (8,192 features, k = 64), LoRA and IA^3 parameter-efficient fine-tuning, ChromBPNet, Ensembl VEP, PhyloP and PhastCons (PlantRegMap), Gemma 4 E4B LLM agent, Baseline gLMs: PlantCAD2, PlantCaduceus, AgroNT, GPN, Evo 2, NTv3, Carbon, MarinDNA, DeepCistrome
Software & environment
PyTorch stack with Mamba2 varlen kernels; training codebase forked from Meta's Lingua; bf16 with fully sharded data parallelism and a compiled model graph; SLURM-scheduled jobs; no software versions or containers stated.
Assay conditions & compute
Nodes of eight NVIDIA B200 GPUs: Botanic1-S 77 h (616 GPU-hours), -M 83 h, -L 202 h, -XL 128 h (4,103 GPU-hours); about 9,300 B200 GPU-hours for the four models, excluding ablations, context extension and evaluation.
Reproducibility gaps
No repository for training or evaluation code is linked; weights, corpus and SAE carry a research-only licence from Living Models; the curation audit behind the 545-variant causal benchmark must be requested from the authors.
Limitations (6)
  • Data-pipeline conclusions rest on two seeds per setting at a single fixed 300M model scale and architecture, judged with an imperfect evaluation suite, so the authors caution they are not final. · stated by the authors
  • Cross-species splice, TIS and TTS scores come from stratified subsamples that are noisy and biased upward (median 95% half-width about 0.065 vs 0.034 for Sbal tasks), giving only a rough model ordering. · stated by the authors
  • The causal-variant benchmark covers only SNPs within a 100 kbp region and assumes the validated variant should be among the least likely, although unlabelled nearby variants may also be functional. · stated by the authors
  • The melon agent experiment repeats runs on one locus, so its recall values describe reliability for that variant rather than a general recovery rate. · stated by the authors
  • Data-mixture decisions were steered with the same downstream suite used for the leaderboard, creating a potential advantage over fixed competitor models that the authors expect to be small. · stated by the authors
  • The 545-variant curation audit is only available on request and all authors are Living Models employees with a declared financial interest, so the headline variant-prioritisation claims cannot yet be independently replicated. · noted in this summary

Read this if you build or apply DNA language models for crops and want a data-curation recipe, a broad plant benchmark, or released models for zero-shot causal-variant ranking.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTerrail '26Yang '26Burda '25Zhai '26Zhai '24Mendoza‐Rev '24
87% semantic matchPlant Biology
Jean Ogier du Terrail, Tanguy Marchand, Vincent Cabeli+3 more·bioRxiv·Mar 4, 2026· 2 citations